A method and system for predicting the low-cycle fatigue life of turbine bladed disks

Abstract

This invention discloses a method and system for predicting the low-cycle fatigue life of turbine bladed disks, relating to the field of aerospace engine technology. The method includes: acquiring a finite element model of the turbine bladed disk and calculation points of the turbine bladed disk under test; sampling the physical data corresponding to the calculation points to obtain an input sample set; determining a training set based on the input sample set and the finite element model of the turbine bladed disk; training a CNN-DNN model based on the training set to obtain an optimal CNN-DNN model; the CNN-DNN model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a flattened layer, a first dense connection layer, a second dense connection layer, and an output layer connected sequentially; based on the optimal CNN-DNN model, determining multiple sets of turbine bladed disk fatigue life data through sampling, and then calculating the reliability of the turbine bladed disk under test to determine the low-cycle fatigue life. This method achieves more efficient and accurate prediction of the low-cycle fatigue life of turbine bladed disks.

Description

Technical Field

[0001] This invention relates to the field of aerospace engine technology, and in particular to a method and system for predicting the low-cycle fatigue life of turbine bladed disks. Background Technology

[0002] Aero engines are a crucial component of the aviation power field, and their high value is determined by their precise manufacturing requirements and high-efficiency performance demands. In particular, the performance of an aero engine depends to a certain extent on the high-temperature performance and strength of the integral bladed disk or rotor of the low-pressure turbine. To reduce engine failure rates and improve disk reliability, it is necessary to improve the fatigue prediction accuracy of integral bladed disks, providing effective support for enhancing and improving their performance.

[0003] Current methods for predicting the fatigue life of aero-engines primarily rely on knowledge graph-based fault diagnosis and prediction platforms, enabling rapid data-driven fault feature extraction and diagnostic methods and processes. However, the diverse and mixed fault vibration signals collected in actual industrial settings exhibit strong non-stationarity, resulting in poor separability. These issues hinder the effective application of traditional data-driven intelligent diagnostic methods developed based on few-fault modes and single-point faults. Deep learning can meet the adaptive feature extraction requirements for aero-equipment fault diagnosis, reducing reliance on expert experience and signal processing techniques while also mitigating the uncertainties introduced by manually designed feature extraction.

[0004] In the field of deep learning, many scholars have used neural networks for research on fault diagnosis and prediction. However, research has found that convolutional neural networks (CNNs) and deep neural networks (DNNs) each have their own advantages and disadvantages. CNNs have accurate data feature extraction capabilities, but their accuracy in data regression modeling is not high; while DNNs have good regression modeling performance, they are not good at data feature extraction. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting the low-cycle fatigue life of turbine bladed disks, which integrates two neural networks, CNN and DNN, to more efficiently and accurately predict the low-cycle fatigue life of turbine bladed disks.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] This invention provides a method for predicting the low-cycle fatigue life of a turbine bladed disk, comprising:

[0008] Based on the physical parameter data of the turbine disk under test, a finite element analysis is performed on the turbine disk under test to obtain the finite element model of the turbine disk and the calculation point of the turbine disk under test; the calculation point is the location of the minimum fatigue life on the turbine disk under test.

[0009] Based on the randomness of the physical parameter data corresponding to the calculation point, physical data sampling is performed to obtain the input sample set;

[0010] Based on the input sample set and the turbine disk finite element model, a finite element simulation output dataset is determined; the finite element simulation output dataset includes multiple simulation output data; the simulation output data is the minimum fatigue life corresponding to the input sample in the input sample set; the input sample set and the finite element simulation output dataset constitute a training set;

[0011] A CNN-DNN model is trained based on the training set to obtain the optimal CNN-DNN model; the CNN-DNN model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a flattened layer, a first dense connection layer, a second dense connection layer, and an output layer connected in sequence.

[0012] Based on the optimal CNN-DNN model, multiple sets of turbine bladed disk fatigue life data are determined by sampling.

[0013] Based on multiple sets of turbine bladed disk fatigue life data, the reliability of the turbine bladed disk under test is calculated, and then the low-cycle fatigue life of the turbine bladed disk under test is determined.

[0014] Optionally, the step of performing finite element analysis on the turbine disk under test based on its physical parameter data to obtain the finite element model of the turbine disk and the calculation points of the turbine disk under test specifically includes:

[0015] The physical parameter data of the turbine bladed disk under test is used as random input variables to establish a finite element model of the turbine bladed disk; the physical parameter data of the turbine bladed disk under test includes the temperature, rotational speed, material parameters and low-cycle fatigue performance parameters of the turbine bladed disk;

[0016] The boundary conditions of the turbine disk finite element model are set based on the coupling effect of temperature load and centrifugal load.

[0017] Based on the boundary conditions and the finite element model of the turbine disk, the calculation points of the turbine disk under test are determined.

[0018] Optionally, based on the randomness of the physical parameter data corresponding to the calculation point, physical data sampling is performed to obtain an input sample set, specifically including:

[0019] Based on the randomness of the physical parameter data corresponding to the calculation points, the Latin hypercube sampling method is used to sample the physical data to obtain the input sample set.

[0020] Optionally, a CNN-DNN model is trained based on the training set to obtain the optimal CNN-DNN model, specifically including:

[0021] The training set is then subjected to normalization and data segmentation processes in sequence to obtain a training dataset and a validation dataset to be used.

[0022] The training dataset to be used is input into the CNN-DNN model to obtain preliminary output data;

[0023] The mean squared error and mean absolute error are calculated based on the initial output data, and the weight parameters of the first dense connection layer and the second dense connection layer in the CNN-DNN model are iteratively adjusted based on the mean squared error and the mean absolute error.

[0024] After the number of iterations reaches the preset number of iterations, the CNN-DNN model after multiple iterations is validated using the validation dataset to be used, and the validation results are obtained.

[0025] When the verification result meets the preset verification conditions, the CNN-DNN model after multiple iterations will be output as the optimal CNN-DNN model.

[0026] If the verification result does not meet the preset verification conditions, the process returns to the step of inputting the training dataset to be used into the CNN-DNN model to obtain preliminary output data.

[0027] Optionally, the calculation formula for each neuron in the first dense connection layer is:

[0028]

[0029] Where p represents the output of the neuron, w r t represents the r-th weight of the neuron connection. r denoted as r-th output of a neuron in a flat layer, b represents the deviation of a single neuron, and a represents the number of connections.

[0030] Optionally, based on the optimal CNN-DNN model, multiple sets of turbine bladed disk fatigue life data are determined by sampling, specifically including:

[0031] Obtain the random input parameter values ​​of the optimal CNN-DNN model;

[0032] Monte Carlo sampling is performed on the random input parameter values ​​to obtain a fatigue life input data set;

[0033] The fatigue life input data set is input into the optimal CNN-DNN model to obtain the fatigue parameter output data set; the output data in the fatigue parameter output data set corresponds one-to-one with the input data in the fatigue life input data set, and the output data in the fatigue parameter output data set and the corresponding input data in the fatigue life input data set constitute a set of turbine bladed disk fatigue life data.

[0034] Optionally, the reliability calculation process for the turbine disk under test is as follows:

[0035] Z = y min -y *

[0036] Where Z represents the limit state function, y * This represents the allowable low-cycle fatigue life value of the turbine bladed disk under test, y. min This represents the actual calculated value, i.e., the fatigue parameter output data of the optimal CNN-DNN model;

[0037]

[0038]

[0039] Where E(Z) represents the mean function; D(Z) represents the variance function, μ = [μ1, μ2…μ n [] represents the mean matrix that the random input parameter values ​​follow; D = [D1, D2, ..., D] n ] represents the variance matrix that the random input parameter values ​​follow, P r This represents the reliability of the turbine disk under test, i.e., the probability that the turbine disk under test will achieve the predetermined power output; μ Z The mean matrix representing the limit state function; D Z φ represents the variance matrix of the limit state function; φ represents the normal distribution function.

[0040] Optionally, the method for predicting the low-cycle fatigue life of the turbine bladed disk further includes:

[0041] Sensitivity analysis is performed on the turbine bladed disk under test based on multiple sets of turbine bladed disk fatigue life data and the reliability of the turbine bladed disk under test.

[0042] To achieve the above objectives, the present invention also provides the following technical solutions:

[0043] A turbine bladed disk low-cycle fatigue life prediction system includes:

[0044] The finite element analysis module is used to perform finite element analysis on the turbine bladed disk under test based on the physical parameter data of the turbine bladed disk under test, so as to obtain the finite element model of the turbine bladed disk and the calculation point of the turbine bladed disk under test; the calculation point is the location of the minimum fatigue life on the turbine bladed disk under test.

[0045] The sampling module is used to perform physical data sampling based on the randomness of the physical parameter data corresponding to the calculation point to obtain the input sample set;

[0046] The training set determination module is used to determine the finite element simulation output dataset based on the input sample set and the turbine bladed disk finite element model; the finite element simulation output dataset includes multiple simulation output data; the simulation output data is the minimum fatigue life corresponding to the input sample in the input sample set; the input sample set and the finite element simulation output dataset constitute the training set;

[0047] The model training module is used to train a CNN-DNN model based on the training set to obtain the optimal CNN-DNN model; the CNN-DNN model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a flattened layer, a first dense connection layer, a second dense connection layer, and an output layer connected in sequence.

[0048] The data set sampling module is used to determine multiple sets of turbine bladed disk fatigue life data sets by sampling based on the optimal CNN-DNN model.

[0049] The life calculation module is used to calculate the reliability of the turbine bladed disk under test based on multiple sets of turbine bladed disk fatigue life data, and then determine the low-cycle fatigue life of the turbine bladed disk under test.

[0050] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0051] This invention provides a method and system for predicting the low-cycle fatigue life of turbine bladed disks. It obtains an input sample set that conforms to actual conditions through finite element analysis and sampling, thus providing high-precision data for subsequent model training and life prediction. Then, fully considering the high-precision feature extraction advantage of convolutional neural networks (CNNs) and the high modeling accuracy advantage of deep neural networks (DNNs) in regression modeling, a CNN-DNN model is constructed, comprising a sequentially connected input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a flattening layer, a first dense connection layer, a second dense connection layer, and an output layer. The first and second convolutional layers are used to extract data features, ensuring the effectiveness of the extracted features. Then, based on the obtained data features, a two-layer dense connection layer is used to perform regression modeling on the DNN to ensure the accuracy of turbine low-cycle fatigue life modeling and probabilistic prediction analysis. By fully utilizing the advantages of CNNs and DNNs, the effectiveness of predicting the low-cycle fatigue life of the turbine bladed disk is guaranteed. Finally, by using the optimal CNN-DNN model obtained through training, multiple sets of turbine bladed disk fatigue life data are efficiently and quickly acquired, and then the reliability of the turbine bladed disk under test is calculated to determine the low-cycle fatigue life of the turbine bladed disk under test. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart illustrating the method for predicting the low-cycle fatigue life of turbine bladed disks according to the present invention.

[0054] Figure 2 This is a schematic diagram illustrating the principle of a one-dimensional convolutional neural network layer.

[0055] Figure 3 This is a schematic diagram illustrating the principle of a deep neural network.

[0056] Figure 4 A scatter plot of predicted and actual values;

[0057] Figure 5 A histogram showing the relative error between predicted and actual values ​​in the range of 0-0.1;

[0058] Figure 6 A histogram showing the relative error between predicted and actual values ​​in the range of 0-0.05;

[0059] Figure 7 A line graph showing the relative error between predicted and actual values;

[0060] Figure 8 This is a schematic diagram of the turbine blade disk low-cycle fatigue life prediction system of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] As a typical component of aero-gas turbine engines, the turbine bladed disk's fatigue life directly affects the reliability and safety of the turbine bladed disk and the entire engine. To address this, this invention proposes a method and system for predicting the low-cycle fatigue life (LCF) of turbine bladed disks. To accurately predict the LCF of turbine bladed disks, an efficient CNN-DNN method is proposed, combining the advantages of Convolutional Neural Networks (CNNs) and Deep Neural Networks (DNNs).

[0063] To make the objectives, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] Example 1

[0065] like Figure 1 As shown, this embodiment provides a method for predicting the low-cycle fatigue life of a turbine bladed disk, including:

[0066] Step 100: Based on the physical parameter data of the turbine disk under test, perform finite element analysis on the turbine disk under test to obtain the finite element model of the turbine disk and the calculation point of the turbine disk under test; the calculation point is the location of the minimum fatigue life on the turbine disk under test.

[0067] Step 100 specifically includes:

[0068] 1) The physical parameter data of the turbine disk under test is used as random input variables to establish a finite element model of the turbine disk; the physical parameter data of the turbine disk under test includes the temperature, rotational speed, material parameters and low-cycle fatigue performance parameters of the turbine disk.

[0069] 2) Based on the coupling effect of temperature load and centrifugal load, the boundary conditions of the turbine disk finite element model are set.

[0070] 3) Based on the boundary conditions and the finite element model of the turbine disk, determine the calculation points of the turbine disk to be tested.

[0071] Step 200: Based on the randomness of the physical parameter data corresponding to the calculation point, physical data sampling is performed to obtain the input sample set; specifically, considering the distribution characteristics of the physical parameter data corresponding to the calculation point, the Latin hypercube sampling method is used to sample the physical data.

[0072] Step 300: Based on the input sample set and the turbine disk finite element model, determine the finite element simulation output dataset; the finite element simulation output dataset includes multiple simulation output data; the simulation output data is the minimum fatigue life corresponding to the input sample in the input sample set; the input sample set and the finite element simulation output dataset constitute the training set. That is, the data in the input sample set is input into the turbine disk finite element model, and the turbine disk finite element model outputs the fatigue life output response at the calculated point, i.e., the finite element simulation output data.

[0073] Step 400: Train a CNN-DNN model based on the training set to obtain the optimal CNN-DNN model; the CNN-DNN model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a flattened layer, a first dense connection layer, a second dense connection layer, and an output layer connected in sequence.

[0074] For the convolutional neural network layer in a CNN-DNN model, i.e., the first convolutional layer or the second convolutional layer:

[0075] Convolutional layers are used to extract features from input vectors. Convolution is a mathematical operation between vectors and filters or kernels. Imagine a one-dimensional filter multiplies the inputs and then sums them to obtain the output, such as... Figure 2 As shown. Where x = (x1, x2, ..., x...) i ,…,x n ) is the input vector, where x s×(i-1)+j It is the s×(i-1)+j-th element of the input vector, where n is the length or number of elements of the input vector, and s represents the span; c = (c1, c2, ..., c j ,…,c k ) is the filter vector, where c j This represents the j-th element, and k represents the length of the filter vector or the number of elements; y = (y1, y2, ..., i ,…,y m ) is the output vector, where y i Let represent the i-th element, and m represent the length of the output vector or the number of elements. The relationship between the output vector length m and the filter vector length k can be expressed as:

[0076]

[0077] For the i-th output response y i The formula for calculating a layer in a one-dimensional convolutional neural network is:

[0078]

[0079] when When the value is not an integer, the length n of the input vector can be updated using a padding method, which includes two methods: (1) padding the input vector with zeros (zero padding) to make the input vector fit the filter; (2) deleting the part of the vector that is not suitable for the filter, which is called effective padding, and only retaining the effective part of the input vector. The present invention uses the effective padding method.

[0080] Regarding the number of filters, this invention uses multiple filters; therefore, the formula for calculating the number of parameters in a one-dimensional convolutional neural network is:

[0081] n p =k×D input ×D output

[0082] Where, n p It is the number of parameters, D input D is the number of filters in the previous one-dimensional convolutional network. output This represents the number of filters in the one-dimensional convolutional network of this layer.

[0083] The number of output variables in a one-dimensional convolutional neural network is calculated using this formula.

[0084] V output =m×D output

[0085] The number of output variables, i.e., the output volume, is V. output .

[0086] For deep neural network layers in CNN-DNN models, namely the first dense connection layer or the second dense connection layer:

[0087] Deep neural networks are hierarchical representations of data. The term "deep" refers to the multiple layers. When data is fed into a neural network model, the network attempts to transform the data by applying specific operations at each layer, representing it in different ways and with different dimensions. Typically, neural networks are simply combinations of linear operations and non-linear functions, such as... Figure 3 As shown in the diagram, circles represent individual neurons, and arrows represent the connection methods of neurons. The diagram shows dense connections. For a single neuron, the output value is the output of each neuron in the previous layer multiplied by its corresponding weight, plus the bias of that neuron.

[0088] Where p represents the output of the neuron, w r t represents the r-th weight of the neuron connection. r Let represent the r-th output of a neuron in a flat layer, b represent the deviation of a single neuron (b is a constant), a represent the number of connections (the number of arrows from the previous layer to this neuron), and ∑ represent the sum.

[0089] In addition, an activation function is needed to increase the complexity and dimensionality of the network. F(·) represents the activation function, and the formula for calculating the output of a neuron after applying the activation function is as follows:

[0090]

[0091] The activation functions (which may differ in each layer) are fixed, but the initial weights and biases are randomized. When the network is trained with input data, the model learns the correct weights and biases and uses a technique called backpropagation to adjust the network accordingly, which changes the network's weights and biases. Once the correct weights and biases are learned, the network will be able to provide meaningful predictions—the values ​​of the network's final layer, the output layer.

[0092] Based on the above convolutional neural network layer and deep neural network layer, step 400 specifically includes:

[0093] 1) The training set is normally processed and data segmented sequentially to obtain the training dataset and the validation dataset to be used.

[0094] 2) Input the training dataset to be used into the CNN-DNN model to obtain preliminary output data.

[0095] 3) Calculate the mean squared error and mean absolute error based on the preliminary output data, and iteratively adjust the weight parameters of the first dense connection layer and the second dense connection layer in the CNN-DNN model based on the mean squared error and the mean absolute error.

[0096] 4) When the number of iterations reaches the preset number of iterations, the CNN-DNN model after multiple iterations is validated using the validation dataset to be used, and the validation results are obtained.

[0097] When the verification result meets the preset verification conditions, the CNN-DNN model after multiple iterations and adjustments will be output as the optimal CNN-DNN model.

[0098] 5) If the verification result does not meet the preset verification conditions, return to the step of inputting the training dataset to be used into the CNN-DNN model to obtain preliminary output data.

[0099] In the training process described above, an early stopping technique is used to prevent overfitting. Overfitting is a phenomenon that occurs during neural network training. When training the model, a portion of the data is also used as a validation dataset. This validation dataset is not used in training but only in the calculation of the loss function and monitoring values. As training progresses, both the training loss and validation loss decrease. However, when the training time approaches a certain point, the validation loss increases instead of decreasing, while the training loss continues to decrease. This indicates that the neural network's memory capacity is greater than that of the training set, and continuing training would make the model too well-suited to the training set and less suitable for samples that the model has never seen before. Therefore, a technique called early stopping is used, employing the validation loss as a monitor to stop training when the validation loss increases.

[0100] Step 500: Based on the optimal CNN-DNN model, multiple sets of turbine bladed disk fatigue life data are determined by sampling.

[0101] Step 500 specifically includes: 1) obtaining random input parameter values ​​of the optimal CNN-DNN model; 2) performing Monte Carlo sampling on the random input parameter values ​​to obtain a fatigue life input data set; 3) inputting the fatigue life input data set into the optimal CNN-DNN model to obtain a fatigue parameter output data set; the output data in the fatigue parameter output data set corresponds one-to-one with the input data in the fatigue life input data set, and the output data in the fatigue parameter output data set and the corresponding input data in the fatigue life input data set constitute a set of turbine bladed disk fatigue life data.

[0102] Step 600: Based on multiple sets of turbine bladed disk fatigue life data sets, calculate the reliability of the turbine bladed disk under test, and then determine the low-cycle fatigue life of the turbine bladed disk under test.

[0103] The calculation process for the reliability of the turbine disk under test is as follows:

[0104] Z = y min -y *

[0105] Where Z represents the limit state function, y * This represents the allowable low-cycle fatigue life value of the turbine bladed disk under test, y. min This represents the actual calculated value, i.e., the fatigue parameter output data of the optimal CNN-DNN model.

[0106] The limit state function Z>0 means that the actual minimum fatigue life of the bladed disk is greater than the allowable value, and the bladed disk is in a safe state; otherwise, the bladed disk structure fails. Assume that the input random variables are independent and follow a mean and variance matrix of μ=[μ1,μ2…μ…]. n ] and D = [D1, D2…D n If the distribution follows a normal distribution, then the following formula applies:

[0107]

[0108]

[0109] Where E(Z) represents the mean function; D(Z) represents the variance function, μ = [μ1, μ2…μ n The mean matrix is ​​the matrix of random input parameter values; D = [D1, D2, ..., D]. n P is the variance matrix that the random input parameter values ​​follow. r This represents the reliability of the turbine disk under test, i.e., the probability that the turbine disk under test will achieve the predetermined power output; μ Z The mean matrix representing the limit state function; D Z φ represents the variance matrix of the limit state function; φ represents the normal distribution function.

[0110] In a specific application, the method for predicting the low-cycle fatigue life of a turbine bladed disk further includes: performing sensitivity analysis on the turbine bladed disk under test based on multiple sets of turbine bladed disk fatigue life data and the reliability of the turbine bladed disk under test.

[0111] Sensitivity is the basis for determining the degree of influence of an input random variable, reflecting how sensitive the input random variable is to the failure (or reliability) of the structural system's response. Sensitivity can be expressed as the reliability with respect to the mean matrix μ of the input random variable. Z Sum of variance matrix D Z The partial derivative of is calculated using the following formula:

[0112]

[0113]

[0114] Where, μ T The above formula can be transformed into: (The formula is the transpose of the mean matrix μ.)

[0115]

[0116] The following analysis uses a turbine blade of a certain type of aero-engine as the research object, and employs the aforementioned method for predicting the low-cycle fatigue life of turbine bladed disks:

[0117] Step (1): Finite Element Analysis

[0118] The turbine blades of a certain type of aero-engine are made of GH4133 high-temperature alloy. The LCF (low-cycle fatigue life) reliability analysis of its integral bladed disk involves multiple variables. The main variables are selected for low-cycle fatigue probability analysis of the integral bladed disk. The selected main variables are as follows: disk temperature T (uncertain due to fluctuations in operating conditions); integral bladed disk velocity ω (reflecting the uncertainty of centrifugal load); density ρ; thermal conductivity; elastic modulus E (reflecting the uncertainty of material parameters); fatigue strength coefficient σ'. f Fatigue ductility coefficient ε' f Fatigue strength index b and fatigue ductility index c.

[0119]

[0120] Where Δε is the total strain; N f It is LCF. Under the complex loads of the engine, considering the influence of average stress, the above formula can be transformed into:

[0121]

[0122] Where, σ m It is the mean stress, fatigue strength coefficient σ' f fatigue ductility coefficient ε' f The fatigue strength index b and fatigue ductility index c reflect the uncertainty of the material parameters in the above formula.

[0123] The selected input random variables are independent and follow a normal distribution. The distribution characteristics of the input random variables are shown in Table 1.

[0124] Table 1 Random Variables and Data Characteristics

[0125]

[0126] In the static analysis of the turbine bladed disk, the stress caused by aerodynamic loads is much smaller than that caused by centrifugal and thermal loads, therefore aerodynamic loads are not considered. Taking into account the coupling effects of temperature and centrifugal loads, appropriate boundary conditions are set, and deterministic analysis is performed. Thermo-structural coupling simulation analysis of the turbine bladed disk was completed on the Workbench 16.0 platform. Since the turbine bladed disk is symmetrical, to reduce the computational load during the analysis, a 1 / 40 scale integral turbine bladed disk is used as the research object. Tetrahedral elements are used to mesh the integral turbine bladed disk, thus obtaining the finite element model of the turbine bladed disk.

[0127] Based on the finite element model of the turbine bladed disk, a thermodynamic analysis of the turbine bladed disk operating at high temperatures is first performed. The heat from the high-temperature gas is transferred to the turbine bladed disk structure through heat conduction and convection. The temperature of the turbine bladed disk gradually increases from the inside to the outside. Empirical formulas are used to simulate the temperature field of the bladed disk, and a thermo-mechanical coupling analysis is conducted. The calculation formulas for the thermo-mechanical coupling analysis are as follows:

[0128]

[0129] Among them, T a Temperature at the leaf root; T b For the blade tip temperature, m = 2 for GH4133B high-temperature alloy.

[0130] The temperature field was applied as a load to the static analysis. Displacement constraints were applied to the inner diameter of the impeller, restricting the translational degrees of freedom along the x, y, and z axes. In addition to rotation along the y and z axes, rotational degrees of freedom were applied to the central hole of the impeller, allowing rotation along the x-axis at a speed of n = 1168 rad / s. The centrifugal load was applied to the impeller structure as a velocity. The average values ​​in Table 1 were used in the established finite element model of the turbine impeller to solve for the thermo-mechanical and structural coupling of the turbine impeller, obtaining the temperature distribution, equivalent stress distribution, and equivalent strain distribution diagrams. According to these three diagrams, the maximum stress and maximum strain of the impeller are located at the blade root, with a maximum stress of 1057.7 MPa and a maximum strain of 8.1427 × 10⁻⁶ MPa. -3 Therefore, the point of maximum strain at the blade root was selected as the research object, and LCF analysis was performed on the bladed disk.

[0131] Furthermore, the deterministic analysis of the fatigue life of the turbine bladed disk comprehensively considers the influence of factors such as structural stress and strain on structural durability. According to the formula... and formula The fatigue life value at the maximum strain point of the impeller structure was calculated, and thus the minimum life of the turbine impeller was found to be 8900.6 cycles.

[0132] Step (2): Neural Network Training

[0133] Using the parameters recorded in Table 1 of step (I) as input random variables, the Latin hypercube sampling method was employed to extract 1000 samples based on the turbine bladed disk finite element model. The input variables and corresponding new data features were then obtained through normalization, as shown in Table 2. The calculation formula for the normalization process is as follows:

[0134]

[0135] Where, x i The input variable vector X = (x1, x2, ..., x) before initialization k The i-th element in ); x'i It is the initialized input variable vector X' = (x1', x2', ..., x...). k The i-th element in X; k is the number of input variables; min(X) represents the smallest element in X; max(X) represents the largest element in X.

[0136] Table 2 Random Variables and Data Characteristics

[0137]

[0138] Out of 1000 samples, 800 samples were selected as training samples for the CNN-DNN model to build the CNN-DNN model, and the remaining 200 samples were used as test samples for the CNN-DNN model to verify the effectiveness of the built CNN-DNN model.

[0139] In this example, the CNN-DNN model uses an eight-layer network with a total of 6779 parameters. The first layer is the input layer with 10 neurons (10 inputs); the second and third layers are convolutional layers with a filter size of 5 and 10 filters, using ReLU as the activation function; the fourth layer is a maxpooling layer with a filter size of 3; the fifth layer is a flattening layer, designed to transform the output of the maxpooling layer into a one-dimensional vector (i.e., the output depth of the flattening layer is 1); the sixth and seventh layers are densely connected layers, each with 64 neurons, using ReLU (Rectified Linear Unit) as the activation function; and the eighth layer is the output layer with only one output neuron. The model is then compiled based on this multi-layered network, using mean squared error (MSE) as the loss function and Adam as the optimization tool for propagation, with both MSE and mean absolute error (MAO) as monitoring values.

[0140] After training begins, the training inputs (corresponding to the input sample set) and training labels (corresponding to the finite element simulation output data) are input into the constructed CNN-DNN model. The batch size is set to 20, and the maximum number of iterations is set to 1000. During training, 20% of the training set is split into a validation set. The validation set is used for calculating the monitoring values ​​and participates in the backpropagation calculation.

[0141] After training, historical plots of mean absolute error (MAO) and mean squared error (MSE) were plotted for both the validation and training sets. The model was then evaluated using the test set, and the MAO and MSE for the test set were calculated. The historical plots showed that the MAO and MSE for both the test and training sets decreased as training progressed. The MAO of the test set was used as a stop-training threshold; therefore, training stopped when the MAO of the test set ceased to improve.

[0142] The trained model was then used to predict the test set, and the prediction results were compared with the labels of the test set. The comparison between the true values ​​and the predicted values ​​is shown in Table 3.

[0143] Table 3 Comparison of Actual and Predicted Values

[0144]

[0145] As shown in Table 3, the predicted values ​​are very close to the true values. However, the test data are out of order because the data is shuffled before being input into the model during training or prediction using a neural network. To observe the difference between the predicted and true values, a scatter plot is created, see... Figure 4 . Figure 4 In the model, when a point is above the center line, it means the predicted value is greater than the actual value; when a point is below the center line, it means the actual value is greater than the predicted value. Regardless of the range of the output value, the predicted value is very close to the actual value.

[0146] Then calculate the relative error between the predicted and actual values ​​and plot a histogram. See Figure 5 and Figure 6 .in Figure 5 A histogram showing the relative error between predicted and actual values ​​in the range of 0-0.1. Figure 6 This is a histogram showing the relative error between predicted and actual values ​​in the range of 0-0.05. Based on... Figure 5 and Figure 6 It can be seen that the relative error between the true value and the predicted value of the test set samples is mostly within 2%, the relative error between the true value and the predicted value of some samples is between 2% and 5%, and only a few samples have a relative error between the true value and the predicted value of more than 5%.

[0147] To determine the degree to which all test set samples deviate from the true values, a line graph showing the relative errors of the true values ​​and the prediction results for all test set samples is plotted with the samples as the x-axis. (See...) Figure 7 .from Figure 7 As can be seen, the relative error between the true and predicted values ​​for most samples is below 2.5%, while only a few samples have a relative error between the true and predicted values ​​that is close to or around 7.5%, indicating that the predictions given by the well-trained CNN-DNN model are very stable.

[0148] Step (3) Reliability Calculation

[0149] 10,000 samples were extracted using the Monte Carlo method and fed into the constructed CNN-DNN model to obtain the corresponding predicted value y. min Based on the reliability calculation process described above, the reliability of the CNN-DNN model output response at different LCF allowable values ​​is calculated, as shown in Table 4.

[0150] Table 4 Reliability of different LCF allowable values

[0151]

[0152] According to Table 4, when the confidence interval is 0.99, the allowable LCF value is 5300 cycles, and the reliability is approximately 0.996. Therefore, considering the safety and reliability of the turbine disk, the allowable LCF value of the disk is set to 5300 cycles.

[0153] Step (IV) Sensitivity Calculation

[0154] To investigate the impact of random parameters on the LCF (Limited Flow Rate), the Monte Carlo method was used to randomly sample 100,000 input random variables. The sample points were normalized and replaced with a trained CNN-DNN model to calculate the output response value. The sensitivity of the turbine disk LCF was calculated according to the sensitivity calculation formula described above, as shown in Table 5.

[0155] Table 5. Sensitivity and Influence Ratio of Random Variables

[0156]

[0157]

[0158] Based on the LCF (Lead Fatigue Capacity) sensitivity of the turbine bladed disk, histograms and proportional distribution maps of the LCF sensitivity distribution were plotted. According to the histograms and proportional distribution maps, and in conjunction with Table 5, the fatigue ductility index c is the main factor affecting the LCF of the turbine bladed disk, accounting for 28% of the total impact. The sensitivity of each variable to the turbine bladed disk is indicated by "-", signifying a negative correlation between the input variable and the LCF. Considering the LCF failure of the turbine bladed disk, the fatigue ductility coefficient c has the greatest impact on the low-cycle fatigue reliability of the turbine bladed disk.

[0159] Step (5): Model Comparison

[0160] The LCF of the turbine blade was modeled and predicted using CNN and DNN respectively, and the relative error between the predicted value and the true value was calculated, as shown in Table 6, and compared with the prediction results of CNN-DNN.

[0161] Table 6. Sensitivity and Influence Ratio of Random Variables

[0162]

[0163] Based on the prediction results of the three methods, plot the relative error between CNN predictions and true values, the relative error between DNN predictions and true values, and the relative error between CNN-DNN predictions and true values.

[0164] According to the relative error graph of CNN predictions and ground truth, using the CNN network layers, there are 15 samples with the most accurate judgment. As the judgment accuracy decreases, the number of samples also decreases. However, many samples have large errors. According to Table 6, the average relative error is 0.0546.

[0165] According to the relative error graph of DNN predicted values ​​and true values, the DNN network layer can identify 10 samples with the highest accuracy. As the accuracy decreases, the number of samples does not decrease much. Although the number of samples with the highest accuracy is less than that of CNN, Table 6 shows that the average relative error is 0.0285.

[0166] According to the relative error graph of CNN-DNN predicted values ​​and true values, using two CNN layers followed by two DNN layers resulted in 14 samples with the most accurate judgments. As the judgment accuracy decreased, there were not many samples with large errors. Table 6 shows that the average relative error is 0.0181. The CNN-DNN model first uses two convolutional layers and then two densely connected layers to extract features from the data and then analyzes the relationship between the simplified input parameters. The resulting average relative error is smaller than that obtained by using CNN or DNN alone.

[0167] 10,000 samples were extracted using the McLeod method and fed into the constructed CNN-DNN model. The predicted value y was obtained from the CNN model and the DNN model. min The reliability of the output response of each model at different LCF allowable values ​​was calculated, as shown in Table 7.

[0168] Table 7 Comparison of Reliability Analysis Results of Three Methods

[0169]

[0170] As shown in Table 7, the reliability calculated by the CNN-DNN model is slightly lower than that of the CNN model and the DNN model. The CNN-DNN model proposed in this paper is more conservative in terms of computational reliability, which is beneficial to improving the security of engineering applications.

[0171] In summary, this invention first employs a double convolutional layer of the CNN to extract data features in the established CNN-DNN model. Then, based on the obtained data features, a double densely connected layer is considered to perform regression modeling on the DNN. Finally, taking the LCF fatigue probability analysis of a turbine bladed disk as an example, a comparison between DNN and CNN shows that the relative error of the CNN-DNN model (0.0181) is significantly smaller than that of CNN (0.0546) and DNN (0.0285), proving that the model has high prediction accuracy. The LCF fatigue probability analysis results of the turbine bladed disk show that when the allowable LCF value is 5300 cycles, the fatigue reliability R is approximately 0.996, and the fatigue ductility index c is the main factor affecting the LCF of the turbine bladed disk, providing a new and effective modeling method for turbine bladed disk LCF prediction.

[0172] Example 2

[0173] like Figure 8 As shown, in order to execute the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, this embodiment provides a turbine bladed disk low-cycle fatigue life prediction system, including:

[0174] The finite element analysis module 101 is used to perform finite element analysis on the turbine bladed disk under test based on the physical parameter data of the turbine bladed disk under test, so as to obtain the finite element model of the turbine bladed disk and the calculation point of the turbine bladed disk under test; the calculation point is the location of the minimum fatigue life on the turbine bladed disk under test.

[0175] The sampling module 201 is used to perform physical data sampling based on the randomness of the physical parameter data corresponding to the calculation point to obtain an input sample set.

[0176] The training set determination module 301 is used to determine the finite element simulation output dataset based on the input sample set and the turbine bladed disk finite element model; the finite element simulation output dataset includes multiple simulation output data; the simulation output data is the minimum fatigue life corresponding to the input sample in the input sample set; the input sample set and the finite element simulation output dataset constitute the training set.

[0177] The model training module 401 is used to train a CNN-DNN model based on the training set to obtain the optimal CNN-DNN model; the CNN-DNN model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a flattened layer, a first dense connection layer, a second dense connection layer, and an output layer connected in sequence.

[0178] The data set sampling module 501 is used to determine multiple sets of turbine bladed disk fatigue life data sets by sampling based on the optimal CNN-DNN model.

[0179] The life calculation module 601 is used to calculate the reliability of the turbine bladed disk under test based on multiple sets of turbine bladed disk fatigue life data sets, and then determine the low-cycle fatigue life of the turbine bladed disk under test.

[0180] Compared with the prior art, the present invention also has the following advantages:

[0181] (1) The CNN-DNN model established in this invention has good prediction performance. This method can accurately predict the LCF of turbine bladed disks without any noise failure samples, and has good robustness. When the CNN-DNN model processes the test set, the maximum sample bias is about 7.5%. The method comparison shows that the CNN-DNN method used has a large number of most accurate samples, a small average error, and the reliability value is calculated slightly conservatively.

[0182] (2) Based on the turbine bladed disk sensitivity analysis, this invention shows that the fatigue ductility index c is the main factor affecting the LCF of the turbine bladed disk, accounting for 28% of the influence, and the fatigue ductility index c is positively correlated with the LCF of the turbine bladed disk. This conclusion provides a basis for the optimized design of the LCF of the turbine bladed disk.

[0183] (3) The CNN-DNN model proposed in this invention has high modeling accuracy and simulation efficiency in predicting the LCF of turbine bladed disks, providing a new method for strength and life analysis of complex mechanical structures.

[0184] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0185] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting low cycle fatigue life of a turbine blisk, characterized by, The method for predicting the low-cycle fatigue life of a turbine bladed disk includes: Based on the physical parameter data of the turbine disk under test, a finite element analysis is performed on the turbine disk under test to obtain the finite element model of the turbine disk and the calculation point of the turbine disk under test; the calculation point is the location of the minimum fatigue life on the turbine disk under test. Based on the randomness of the physical parameter data corresponding to the calculation point, physical data sampling is performed to obtain the input sample set; Based on the input sample set and the turbine disk finite element model, a finite element simulation output dataset is determined; the finite element simulation output dataset includes multiple simulation output data; the simulation output data is the minimum fatigue life corresponding to the input sample in the input sample set; the input sample set and the finite element simulation output dataset constitute a training set; A CNN-DNN model is trained based on the training set to obtain the optimal CNN-DNN model; the CNN-DNN model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a flattened layer, a first dense connection layer, a second dense connection layer, and an output layer connected in sequence. Based on the optimal CNN-DNN model, multiple sets of turbine bladed disk fatigue life data are determined by sampling. Based on multiple sets of turbine bladed disk fatigue life data, the reliability of the turbine bladed disk under test is calculated, and then the low-cycle fatigue life of the turbine bladed disk under test is determined.

2. The method of predicting low cycle fatigue life of a turbine bladed disk according to claim 1, wherein, The step of performing finite element analysis on the turbine disk under test based on its physical parameter data to obtain the finite element model of the turbine disk and the calculation points of the turbine disk under test specifically includes: The physical parameter data of the turbine bladed disk under test is used as random input variables to establish a finite element model of the turbine bladed disk; the physical parameter data of the turbine bladed disk under test includes the temperature, rotational speed, material parameters and low-cycle fatigue performance parameters of the turbine bladed disk; The boundary conditions of the turbine disk finite element model are set based on the coupling effect of temperature load and centrifugal load. Based on the boundary conditions and the finite element model of the turbine disk, the calculation points of the turbine disk under test are determined.

3. The method of claim 1, wherein, Based on the randomness of the physical parameter data corresponding to the calculation points, physical data sampling is performed to obtain the input sample set, specifically including: Based on the randomness of the physical parameter data corresponding to the calculation points, the Latin hypercube sampling method is used to sample the physical data to obtain the input sample set.

4. The method of claim 1, wherein, The CNN-DNN model is trained based on the training set to obtain the optimal CNN-DNN model, specifically including: The training set is then subjected to normalization and data segmentation processes in sequence to obtain a training dataset and a validation dataset to be used. The training dataset to be used is input into the CNN-DNN model to obtain preliminary output data; The mean squared error and mean absolute error are calculated based on the initial output data, and the weight parameters of the first dense connection layer and the second dense connection layer in the CNN-DNN model are iteratively adjusted based on the mean squared error and the mean absolute error. After the number of iterations reaches the preset number of iterations, the CNN-DNN model after multiple iterations is validated using the validation dataset to be used, and the validation results are obtained. When the verification result meets the preset verification conditions, the CNN-DNN model after multiple iterations will be output as the optimal CNN-DNN model. If the verification result does not meet the preset verification conditions, the process returns to the step of inputting the training dataset to be used into the CNN-DNN model to obtain preliminary output data.

5. The method of claim 1, wherein, The calculation formula for each neuron in the first densely connected layer is as follows: where p represents the output of the neuron, w r represents the rth weight of the neuron connection, t r represents the rth output of the neuron in the flat layer, b represents the bias of a single neuron, and a represents the number of connections.

6. The method of claim 1, wherein, Based on the optimal CNN-DNN model, multiple sets of turbine bladed disk fatigue life data are determined through sampling, specifically including: Obtain the random input parameter values ​​of the optimal CNN-DNN model; Monte Carlo sampling is performed on the random input parameter values ​​to obtain a fatigue life input data set; The fatigue life input data set is input into the optimal CNN-DNN model to obtain the fatigue parameter output data set; the output data in the fatigue parameter output data set corresponds one-to-one with the input data in the fatigue life input data set, and the output data in the fatigue parameter output data set and the corresponding input data in the fatigue life input data set constitute a set of turbine bladed disk fatigue life data.

7. The method of predicting the low cycle fatigue life of a turbine bladed disk according to claim 6, wherein, The calculation process for the reliability of the turbine disk under test is as follows: Z = y min - y * wherein Z represents a limit state function, y * represents a low-cycle fatigue life allowable value of the turbine blade to be measured, y min represents an actual calculation value, i.e., a fatigue parameter output data output by the optimal CNN-DNN model wherein E(Z) represents a mean function; D(Z) represents a variance function, μ = [μ1, μ2…μ n ] represents a mean matrix to which the random input parameter values are subjected; D = [D1, D2…D n ] represents a variance matrix to which the random input parameter values are subjected, P r represents a reliability of the turbine blade to be tested, i.e. a probability that the turbine blade to be tested completes a predetermined power; μ Z represents a mean matrix of the limit state function; D Z represents a variance matrix of the limit state function; and φ represents a normal distribution function.

8. The method for predicting the low-cycle fatigue life of a turbine bladed disk according to claim 1, characterized in that, The method for predicting the low-cycle fatigue life of turbine bladed disks also includes: Sensitivity analysis is performed on the turbine bladed disk under test based on multiple sets of turbine bladed disk fatigue life data and the reliability of the turbine bladed disk under test.

9. A turbine bladed disk low-cycle fatigue life prediction system, characterized in that, The turbine bladed disk low-cycle fatigue life prediction system includes: The finite element analysis module is used to perform finite element analysis on the turbine bladed disk under test based on the physical parameter data of the turbine bladed disk under test, so as to obtain the finite element model of the turbine bladed disk and the calculation point of the turbine bladed disk under test; the calculation point is the location of the minimum fatigue life on the turbine bladed disk under test. The sampling module is used to perform physical data sampling based on the randomness of the physical parameter data corresponding to the calculation point to obtain the input sample set; The training set determination module is used to determine the finite element simulation output dataset based on the input sample set and the turbine bladed disk finite element model; the finite element simulation output dataset includes multiple simulation output data; the simulation output data is the minimum fatigue life corresponding to the input sample in the input sample set; the input sample set and the finite element simulation output dataset constitute the training set; The model training module is used to train a CNN-DNN model based on the training set to obtain the optimal CNN-DNN model; the CNN-DNN model includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a flattened layer, a first dense connection layer, a second dense connection layer, and an output layer connected in sequence. The data set sampling module is used to determine multiple sets of turbine bladed disk fatigue life data sets by sampling based on the optimal CNN-DNN model. The life calculation module is used to calculate the reliability of the turbine bladed disk under test based on multiple sets of turbine bladed disk fatigue life data, and then determine the low-cycle fatigue life of the turbine bladed disk under test.

Images